On large-deviation efficiency in statistical inference (Q1265761)

The authors give a unified treatment of large-deviations (LD) efficiency in statistical inference in consideration of weak convergence theory. The concept of the LD principle (LDP) for a sequence of statistical experiments is defined both for the dominated case and for the general case. It is explained using the Gaussian shift model, the i.i.d. model, the ``signal plus white noise'' model, the regression model with Gaussian and non-Gaussian errors, etc. Some properties of the LDP for statistical experiments are discussed and a sufficient condition for it is given. It is also shown that if a sequence of statistical experiments obeys the LDP, then there is an asymptotic lower bound for appropriately defined minimax risks, and it is sharp under a strengthened version of the LDP. The authors give a method of obtaining nearly LD efficient decisions in the sense that their DL asymptotic risk is arbitrarily close to the lower bound. Further, the above results are applied to the cases of testing hypothesis and estimation, and previous models.